Haishan Zhu

Title: Haishan Zhu: Innovator in Neural Network Training

Introduction

Haishan Zhu is a prominent inventor based in Bellevue, WA (US). He has made significant contributions to the field of neural networks, particularly in optimizing memory consumption and processor utilization.

Latest Patents

Haishan Zhu holds two patents, with his latest focusing on neural network training with decreased memory consumption and processor utilization. This innovation involves bounding box quantization, which reduces the number of bits used to express numerical values before matrix multiplication. This process effectively lowers both memory consumption and processor utilization. Additionally, stochastic rounding is employed to maintain sufficient precision, allowing weight values to be stored in reduced-precision formats without the need for separate full-precision storage. Other rounding mechanisms, such as rounding to the nearest value, can also be utilized for exchanging weight values in reduced-precision formats while retaining full-precision formats for future updates. To facilitate this conversion, reduced-precision formats like brain floating-point format are utilized.

Career Highlights

Haishan Zhu is currently associated with Microsoft Technology Licensing, LLC, where he continues to innovate in the realm of technology and artificial intelligence. His work has garnered attention for its practical applications in enhancing the efficiency of neural networks.

Collaborations

Some of his notable coworkers include Taesik Na and Daniel Lo, who contribute to the collaborative environment that fosters innovation at Microsoft.

Conclusion

Haishan Zhu's work in neural network training exemplifies the intersection of technology and efficiency. His patents reflect a commitment to advancing the field while addressing critical challenges in memory and processing power.